Comparative Analysis of Deep Learning Models for Olive Detection on
the Branch
ERHAN KAHYA1, YASIN ASLAN2
1Department of Electronic and Automation
Tekirdağ Namık Kemal University
Tekirdağ,
TURKEY
2Freelance Senior Software Developer
Tekirdağ,
TURKEY
Abstract: - The future of deep learning integration in agriculture holds great potential for advancing sustainable
agricultural practices, precision agriculture and improved decision-making. With the rapid development of
image processing and artificial intelligence technologies in recent years, deep learning has begun to play a
major role in identifying agricultural pests and optimizing agricultural product marketing. However, there are
challenges related to data quality, model scalability, and geographical limitations for widespread adoption of
deep learning in agriculture. This study on Olive was conducted to improve the quality of the data set and to
ensure more reliable training of object detection models. According to the result of the training process of
YOLOv7 used in the study, it was concluded that it was characterized by decreasing loss values and showed an
increase in the model's ability to detect objects correctly. It was observed that the other model, YOLOv8l, had a
more effective learning capacity and a tendency to learn faster. The performance of both models was evaluated
with various metrics, and it was determined that YOLOv8l had higher Precision, Recall, and mAP values. It
was emphasized that YOLOv8l showed high performance even in low epoch numbers and can be preferred
especially in cases where time and computational resources were limited. It was determined that YOLOv7
made detections in a wide confidence range, but had difficulty in detections with low confidence scores. It was
observed that YOLOv8l made more stable and reliable detections with higher confidence scores. The metric
data of the "YOLOv8l" model was found to be higher compared to other models. The F1 score of the YOLOv5l
model was 92.337%, precision 96.568%, recall %88,462,mAP@0.5:0.65 value gave the highest score with
94.608%. This research on deep learning-based object detection models indicated that YOLOv8l showed
superior performance compared to YOLOv7 and was a more reliable option for agricultural applications.
Key-Words: - Deep learning, YOLOv8, YOLOv7, identification, analysis, classification.
Received: August 21, 2023. Revised: November 15, 2023. Accepted: December 11, 2023. Published: December 31, 2023.
1 Introduction
Deep learning has had a major impact on the
agricultural sector and has started to be used in
many areas. In agricultural applications, deep
learning is used in crop management, water
management, soil management, livestock
management, and classification. Deep learning is an
artificial intelligence method that uses multi-layered
artificial neural networks and can be used in areas
such as soil analysis, disease, and pest detection in
agriculture. Deep learning is used to increase
agricultural productivity and improve production. In
particular, the ability of deep learning to deal with
large data sets shows that it is highly effective in
agricultural applications, [1]. Deep learning is used
in a wide variety of areas in agricultural
applications. The successes achieved with deep
learning methods in the agricultural image
recognition competition have revealed the potential
in this field, [2]. Deep learning is also used in
processing agricultural sensor data and forecasting
agricultural production, [3]. Deep learning plays an
important role in issues such as sustainability,
productivity, reduction of input costs, food safety,
and environmental sustainability in agricultural
applications, [4]. The use of these technologies in
agricultural applications enables the development of
smarter and sustainable agricultural practices, [5].
Deep learning is used in areas such as agricultural
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
338
Volume 22, 2023
data analytics, disease diagnosis, and production
forecasting to increase the automation and
efficiency of agricultural processes, [6]. Deep
learning models are also used for decision support
systems in agricultural production, [7]. In this
context, it is seen that deep learning techniques are
used in a wide range of agricultural applications and
provide a significant transformation in the
agricultural sector. The use of these techniques
contributes to making agricultural processes more
efficient, sustainable, and smart.
2 Material and Method
2.1 Material
Olive is the fruit of the olive tree and is usually used
for olive oil production or direct consumption.
There are two main types. The first is table olives,
which are used for direct consumption and are
usually pickled in water or stored in oil. The other is
small, bitter-pressed olives that are generally used
for olive oil production. Olive trees generally grow
in regions specific to the Mediterranean climate, and
the harvest time of this fruit varies depending on the
geographical location and olive type. Olives
generally ripen in autumn and are harvested during
this period. The olive harvesting process is a very
important stage in olive oil production as it
significantly affects the quality of the final product.
Olive harvesting can be performed using manual or
mechanical methods, and each approach has specific
effects. Manual harvesting involves manually
picking olives from trees or collecting olives that
have fallen to the ground and is usually done by
farm workers, [8]. On the other hand, mechanical
harvesting uses technological methods such as
shaking trees to dislodge olives. This method may
be more efficient, but can also cause potential
damage to the fruit, [9]. The labor aspect of olive
harvesting is also an important issue. This is
because hand picking of olives involves the
participation of farm laborers, [10]. This emphasizes
the social and occupational health dimensions
associated with olive harvesting and emphasizes the
need for fair labor practices and worker welfare in
agricultural settings.
2.2 Method
While preparing the data set of the olive fruit, which
was targeted for object detection and analysis within
the scope of the study, 140 images taken from the
producer olive field in Tekirdağ Naip Village and
various internet sources were used. The parts
containing each olive image within the images were
marked with a bounding box area. Sample images
taken from the producer's field are shown in Figure
1 (a-b).
(a)
(b)
Fig. 1: a-b: Examples of educational images
2.2.1 Labeling
For an object detection model to be able to train on
a dataset, the objects to be detected must be
labeled/signed in the dataset to be trained. For this
reason, the parts containing the olive image in each
of the 140 images should be marked with the
bounding box area and assigned to the "olive" class,
which is the object class it belongs to. There are
many programs, websites, and tools available in the
open-source communities for image labeling.
RoboFlow is a versatile platform that offers a range
of functions for robotics and artificial intelligence
applications. It enables the development and
training of machine learning models by providing
users with access to public datasets and the ability to
send private data, [1]. The platform also supports
pre-processing of images and their labels, as well as
the creation of bounding boxes for tasks such as
object detection and classification, [10], [11].
RoboFlow enables end users to control robots by
developing flowchart-like programs consisting of
visual markers and making them accessible to
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
339
Volume 22, 2023
people without extensive programming expertise,
[12]. Roboflow offers a flow-based visual
programming language specifically designed for
engineering robot motion and manipulation tasks by
meeting the needs of both inexperienced and
experienced programmers, [13]. This marking and
labeling process is easily done through the graphical
user interface of the site. The Label screen is shown
in Figure 2. a-b.
(a)
(b)
Fig. 2: a-b: Label display
2.2.2 Training Model Selection
In the study we carried out, the YOLOv7 and
YOLOv8 families, developed as an open source of
the YOLO model family developed by the CNN
method, were preferred. The YOLOv7 architecture,
an extension of YOLOv4, has been developed with
several architectural reforms, including the
Extended Efficient Layer Aggregation Network (E-
ELAN) and free trainable tools, [14]. YOLOv7 is
the newest variant of the YOLO object detector
family, known for its high accuracy and fast
implementation, [15]. The YOLOv7-Pose model,
based on YOLO-Pose and YOLOv7, uses E-LAN
and E-ELAN architecture for high accuracy and
fast, [16]. The YOLOv7 architecture includes a
computation block called E-ELAN, which allows
better learning within the framework, [17].
YOLOv8, the latest version of the YOLO series,
has attracted attention with its advances in real-time
object detection. It is designed to provide high
accuracy while maintaining real-time performance
by combining the strengths of various real-time
object detectors, [18]. The YOLOv8 architecture
offers five different size options that meet various
application needs, from nano-size to extra-large
size, [19]. Building on the strengths of its previous
versions, YOLOv8 has emerged as a faster and
more capable option and positions itself as a
superior alternative to existing models, [20]. The
inclusion of object detection skills in the YOLOv8
framework increases its potential for robotic
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
340
Volume 22, 2023
cognitive architectures by contributing to perceptual
fixation capabilities, [21]. YOLOv7 and YOLOv8
versions used in the study contain models within
themselves. YOLOv7, YOLOv8n/s/m, and l (nano-
small-medium and large) models were preferred for
deep learning training.
2.2.3 Initiation of Training
To start the training of the model that will perform
olive detection, the location of the YOLOv7/v8
model on the computer was visited. In the next step,
a Python editor was used to control and run the
train.py program located in the main directory that
manages the YOLOv7/v8 training. This Python
program can be customized with various parameters
and made suitable for specific training for Olive
detection.
Within the project, the parameters and regulations in
the code written below were preferred for olive
fruit.
YOLOv7
!python train.py --device 0 --batch 20 --epochs
100 --data {dataset.location}/data.yaml
--weights 'yolov7_training.pt'
--device: This parameter indicates the device on
which the training will be performed. The value "0"
is generally used for the first graph processing unit
(GPU).
--batch: The number of data point packets to be
used by the display card at a time while training the
model.
--epochs: The number of times all training data is
shown to the trained network and the weights are
updated while training the model
--data: It is a configuration file that defines the
training data of the model. This file contains the
data path, class labels, and other data characteristics.
--weights: Indicates the initial weights to be used
during training of the model. This is important in
transfer learning situations. The file
'yolov7_training.pt' contains the pre-trained weights.
If this parameter is not indicated, the model usually
starts training with random weights.
YOLOV8
!yolo task=detect mode=train model=yolov8n.pt
data={dataset.location}/data.yaml epochs=50
imgsz=640 plots=True
!yolo task=detect mode=train model=yolov8s.pt
data={dataset.location}/data.yaml epochs=50
imgsz=640 plots=True
!yolo task=detect mode=train model=yolov8m.pt
data={dataset.location}/data.yaml epochs=50
imgsz=640 plots=True
!yolo task=detect mode=train model=yolov8l.pt
data={dataset.location}/data.yaml epochs=50
imgsz=640 plots=True
--task: Determines which task the training will
perform. “detect” refers to the object detection task.
--mode: Indicates that the model will be run in
training mode. YOLO can be run in different modes,
for example, “train” can be used for training, and
“test” can be used for testing or validation.
--model: Indicates the path and name of the model
to be used during training.
--imgsz: The pixel size at which the images to be
trained will be reduced by the YOLOv8 model.
--plots: Refers to graph drawings created during
training. These graphs are used to monitor the
performance of the model.
As a result of running these code lines correctly, the
training process of the model started. The program
first checks the YOLOv7/v8 files and checks for any
update status. Then, the training process is carried
out during the determined number of cycles (epoch).
Examination of the results of YOLOv7 and
YOLOv8l algorithms according to error matrix
metrics is shown in Figure 3, Figure 4, Figure 5,
Figure 6, Figure 7, Figure 8 respectively.
(Image 1, Size: 640x640, Batch: 20 Epoch: 100,
Algorithm: YOLOv7)
(Image 2, Size: 640x640, Epoch: 50, Algorithm:
YOLOv8l)
Fig. 3: F1 Score display
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
341
Volume 22, 2023
According to the analysis of the graphs in
Figure 3, the F1 Score result analysis is as
follows.The F1-Confidence curve of the YOLOv7
algorithm at 640x640 resolution and after 100
epochs of training are shown in Image 1. From the
graph, it was seen that the model achieved a high F1
score for the 'olive' class. The maximum F1 score
(about 0.93) was obtained at a confidence threshold
of 0.498. The curve rises at lower confidence
thresholds. After reaching the optimum point, it
decreased with increasing confidence threshold.
This indicated that the model provided a good
balance between the accuracy and scope of the
detections within a given confidence interval.
However, it was understood that after the
confidence threshold of 0.6, the F1 score decreased
rapidly. It was understood that high threshold values
can increase false negatives and the sensitivity of
the model decreased.
The F1-Confidence curve of the YOLOv8l
algorithm obtained at 640x640 resolution and after
50 epochs of training is shown in Image 2. The
model achieved an almost constant high F1 score
(approximately 0.92) for the 'olive' class and this
score reached to maximum at a confidence threshold
of 0.466. The curve is almost a straight line over a
wide confidence interval. This indicated that the
model provided high accuracy and sensitivity even
over a wide confidence interval. The confidence
threshold of the curve decreased sharply at around
0.8, and from this point onwards, it was understood
that the model significantly increased the false-
negative rate and the sensitivity decreased.
In both graphs, the F1 scores of all classes were
also indicated to evaluate the overall performance of
the model. For YOLOv7, the F1 score of all classes
was 0.93, while it was shown as 0.92 for YOLOv8l.
These values indicated that both models showed
high performance and gave good results over a wide
threshold of confidence. For both models, it was
understood that the best performance was achieved
at medium confidence thresholds. It was observed
that higher or lower threshold values negatively
affected performance.
According to the analysis of the graphs in
Figure 4, the Precision result analysis is as
follows.The accuracy-confidence curve of the
YOLOv7 algorithm for the 'olive' class is shown in
Image 1. The curve observed in the image started
with low accuracy at low confidence thresholds and
increased as the confidence threshold increased.
This indicated that the model showed improvement
in the accuracy of detections within a certain
confidence interval. The accuracy reached a
maximum (1.00) at a confidence threshold value of
0.641, indicating that all of the model's detections at
this threshold were correct. However, accuracy
decreased rapidly as the confidence threshold
increased further. This suggests that at very high
confidence thresholds the model increased false
negatives and therefore made fewer true positive
detections.
(Image 1, Size: 640x640, Batch: 20 Epoch:
100,Algorithm: YOLOv7)
(Image 2, Size: 640x640, Epoch: 50, Algorithm:
YOLOv8l)
Fig. 4:Precision display
The truth-confidence curve for the 'olive' class
of the YOLOv8l algorithm is shown in Image 2. The
curve rose rapidly at low confidence thresholds and
reached maximum accuracy (1.00) at a confidence
threshold of 0.511. It was observed that the curve
almost flattened towards the high confidence
thresholds and remained at this level. This showed
that the YOLOv8l algorithm achieved high
accuracy, especially at intermediate confidence
thresholds, and that the model made true positive
detections even above this threshold, keeping the
number of false positives minimum.
The maximum accuracy values for all classes
for both models are indicated in the graph and are
shown as 1.00 in both cases. This showed that the
model reached the maximum accuracy value when
tested on all classes.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
342
Volume 22, 2023
(Image 1, Size: 640x640, Batch: 20 Epoch: 100,
Algorithm: YOLOv7)
(Image 2, Size: 640x640, Epoch: 50, Algorithm:
YOLOv8l)
Fig. 5:Recall display
According to the analysis of the graphs in
Figure 5, the Recall result analysis is as follows.
The sensitivity-confidence curve of the YOLOv7
algorithm for the 'olive' class is shown in Image 1.
The sensitivity started at 1.00 (perfect sensitivity) at
low confidence thresholds and decreased as the
confidence threshold increased. This indicated that
the model was able to detect almost all true
positives at low confidence thresholds. However, it
was seen that as the confidence threshold increased,
the model missed some true positive detections and
the sensitivity decreased. In the graph, the
confidence threshold decreased sharply to around
0.8. It was understood that higher threshold values
significantly reduced the sensitivity by increasing
the false negative rate.
The sensitivity-confidence curve of the
YOLOv8l algorithm for the 'olive' class is shown in
Image 2. Similarly, the model showed high
sensitivity at low confidence thresholds. The
sensitivity started to decrease rapidly at a relatively
low confidence threshold (around 0.2) and
continued to decrease as the confidence threshold
increased. In this model, sensitivity decreased
significantly at high confidence thresholds. This
indicated that the model had a high false negative
rate at high confidence thresholds.
In both graphs, the sensitivity scores of all
classes were indicated to evaluate the overall
performance of the model, and in both cases,
excellent sensitivity (1.00) was achieved at the
lowest confidence threshold. This indicated that
both models were able to detect all true positives at
low confidence thresholds.
(Image 1, Size: 640x640, Batch: 20 Epoch: 100,
Algorithm: YOLOv7)
(Image 2, Size: 640x640, Epoch: 50, Algorithm:
YOLOv8l)
Fig. 6:Precision-Recall display
According to the analysis of the graphs in
Figure 6, the Precision-Recall result analysis is as
follows.In the Precision-Recall curve showing the
performance of the YOLOv7 algorithm in Image 1,
it was seen that a very high precision value was
achieved for the 'olive' class. The graph shows that
the model achieved high accuracy over a wide range
of precision (recall). The mAP@0.5 value of the
model for the 'olive' class was found as 0.947,
indicating that the model detected objects with high
precision.
In the Image 2 graph, it was seen that in the
Precision-Recall curve of the YOLOv8l algorithm,
almost perfect accuracy was obtained for the 'olive'
class. The curve in the graph showed that the model
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
343
Volume 22, 2023
detected true positives with very high accuracy and
the value of mAP@0.5 was 0.981. It was understood
that the classification performance of the model was
outstanding and that it detected very few false
positives.
In both graphs, it was seen that there was a
sharp decrease in accuracy after a certain sensitivity
value. This meant that the model detected all
positive examples, but this caused some false
positives. The graphs showed the overall
performance of all classes while maintaining high
accuracy and sensitivity values for the 'olive' class.
The high mAP@0.5 values of both models indicated
that they performed excellently in object detection.
The graphs showed the overall performance of the
model, as well as how balanced its performance was
for different classes. While both algorithms offered
high accuracy and sensitivity, it was remarkable that
YOLOv8l achieved a very high value of mAP@0.5
even at a lower epoch number. This indicated that
YOLOv8l may be more effective than YOLOv7 in
certain situations.
(Image 1, Size: 640x640, Batch: 20 Epoch: 100,
Algorithm: YOLOv7)
(Image 2, Size: 640x640, Epoch: 50, Algorithm:
YOLOv8l)
Fig. 7: Confusion Matrix display
According to the analysis of the graphs in
Figure 7, the Confusion Matrix result analysis is as
follows.Image 1 showed that the YOLOv7 model
performed almost perfectly for the "olive" class. The
matrix showed a high TP ratio (1.0) for the "olive"
class and a high TN ratio for the "background"
class. This meant that the model detected the "olive"
class almost without error and made almost no false
positive predictions.
In Image 2, the confusion matrix of the
YOLOv8l model again showed a high TP rate for
the "olive" class (0.89), but there was an FP ratio
(0.08) indicating that this model also misclassified
some of the "background" class as "olive". This
meant that although the YOLOv8l model run with a
slightly lower accuracy rate, it still exhibited a quite
high accuracy.
Both matrices showed the high sensitivity of the
models in detecting the "olive" class and the overall
low false positive rate. However, while the
YOLOv7 model almost completely avoided false
positives in the "background" class, the YOLOv8l
model was somewhat weak in this regard. This
indicated that YOLOv7 was able to discriminate
better than YOLOv8l in this particular case.
(Image 1, Size: 640x640, Batch: 20, Epoch: 100,
Algorithm: YOLOv7)
(Image 2, Size: 640x640, Epoch: 50, Algorithm:
YOLOv8l)
Fig. 8: Loss Function display
According to the analysis of the graphs in
Figure 8, the Loss Function result analysis is as
follows.The graphs showed how various loss
functions and performance metrics changed during
the training process of the YOLOv7 and YOLOv8l
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
344
Volume 22, 2023
algorithms. Such graphs are used to understand the
learning process and performance of the model on
training and validation sets.
Image 1 Analysis (YOLOv7)
This series of graphs shows the training process of
the YOLOv7 model trained at 640x640 resolution,
with 20 batch sizes and 100 epochs.
Box Loss: It measures the availability of the
bounding boxes predicted by the model to
the real boxes. A downward trend indicates
that the model starts to predict bounding
boxes more accurately over time.
Objectness Loss: It measures how
accurately the model predicts the presence
of an object. A decrease indicates that the
model better discriminates object presence
over time.
Classification Loss: It measures the
classification accuracy of the model. A
decrease indicates that the model classifies
classes more accurately over time.
Precision ve Recall: It shows the accuracy
(how accurate positives are detected) and
sensitivity (how many of all positives are
detected) values of the model. An increase
in both values indicates an improvement in
the model’s performance.
mAP@0.5 ve mAP@0.5:0.95: It measures
the average accuracy rates of the model. In
general, an increasing trend indicates that
the model can generalize well over all
classes.
Image 2 Analysis (YOLOv8l)
The second series of graphs shows the training
process of the YOLOv8l model trained at 640x640
resolution and 50 epochs.
Train Box/Class/DFL Loss: Bounding
box, classification, and directional focus
loss on the training set of the model.
Downward trends indicate that the model
shows improvement during the training
process.
Val Box/Class/DFL Loss: The losses of the
model on the validation set. Decreases
indicate that the model also shows
improvement in the validation set and does
not make overfitting.
Precision/Recall: Accuracy and precision
of the model on the training set. Stable and
high values indicate the reliability of the
model.
Metrics/mAP50/B: Average accuracy rates
of the model. High mAP values indicate that
the model makes consistent detections over
a wide range.
In both images, it was seen that the performance
and loss functions of the model in the training
process tended to decrease. It was understood that
YOLOv7 had a more stable training process with a
longer training time (100 epochs), while YOLOv8l
showed a high performance in a shorter time (50
epochs). It was observed that both models had high
mAP values and generally showed high accuracy
and sensitivity.
3 Research Results
The training process of YOLOv7 was characterized
by generally decreasing loss values. The decrease in
Box, Objectness, and Classification Loss values
indicated an increase in the model's ability to
correctly detect objects within bounding boxes, and
correctly identify and classify object presence. The
increase in the Precision and Recall metrics
indicated that the accuracy and sensitivity of the
model's detections increased. The metrics mAP@0.5
and mAP@0.5:0.95 indicated the reliability and
interclass consistency of the model with high values.
The loss graphs for YOLOv8l showed a similar
trend of improvement, but higher mAP values were
obtained at a shorter number of epochs. This
indicated that the learning capacity and speed of
YOLOv8l was more efficient than YOLOv7.
Precision and Recall values showed that the model
was highly capable of accurately detecting and
classifying objects.
The confusion matrices showed that YOLOv7
showed almost a perfect performance in
classification, while YOLOv8l ran with a certain
margin of error. The Precision-Recall curves
showed that both models had high Precision and
Recall values, but the YOLOv8l performed nearly
perfectly even at a low epoch count.
Both models exhibited high performance, but it
was remarkable that YOLOv8l had a tendency to
learn faster and achieved high mAP values with
fewer epochs. This indicated that YOLOv8l may be
preferable, especially when time and computational
resources are limited. Despite the longer training
time, YOLOv7 showed a more stable performance
during the learning process. Both models provided
high accuracy and precision, but YOLOv8l provided
a slight advantage in classification.
Extensive analysis of training processes and
performance metrics showed that both models were
suitable for object detection. However, the
performance of YOLOv8l showed that it was more
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
345
Volume 22, 2023
efficient, especially in terms of training time and
computational resources.
Figure 9 shows the test result screen, Figure 10
shows the "Validation Batch" prediction markings
resulting from the training of YOLOv7 and Figure
11 shows the YOLOv8l models.
(Image 1, Size: 640x640, Batch: 20, Epoch: 100,
Algorithm: YOLOv7)
(Image 2,
Size: 640x640, Epoch: 50, Algorithm: YOLOv8l)
Fig. 9: Test result screen
Fig. 10: Size: 640x640, Batch: 20, Epoch: 100,
Algorithm: "Validation Batch" prediction markings
resulting from the training of the YOLOv7 model
Fig. 11:Size: 640x640, Epoch: 50, Algorithm:
"Validation Batch" prediction markings resulting
from the training of the YOLOv8l model
The training of YOLOv7 and YOLOv8l models
with different epoch numbers affects the learning
dynamics. YOLOv7 was trained for 100 epochs and
YOLOv8l was trained for 50 epochs. During the
training process, it was observed that the loss values
in the YOLOv7 model showed a slower decrease,
whereas the YOLOv8l model showed a faster and
more stable decrease trend. This indicates that the
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
346
Volume 22, 2023
learning capacity and speed of YOLOv8l is more
effective than YOLOv7. Furthermore, when the
performances of both models on the validation data
are analyzed, it is concluded that YOLOv8l shows
high accuracy and sensitivity despite the lower
number of epochs, while YOLOv7 is not as
effective despite the longer training time.
The analysis of the hyperparameters (e.g.
learning rate, momentum, weight reduction)
specified in the model's configuration files
significantly affects the performance of each model.
The different sets of hyperparameters used in the
YOLOv7 and YOLOv8l models directly affected
the training processes and results of the models. For
example, the higher learning rate applied in the
YOLOv8l model enabled the model to learn faster
and generalize more effectively. In contrast, the
lower learning rate in the YOLOv7 model caused
the model to learn more slowly but more carefully.
These differences were reflected in the object
detection performances as well as the training
processes of the models.
Analyzing the performance of both models
under different lighting conditions and background
scenarios is an important factor in evaluating the
applicability of the models in real-world
applications. In the tests, the YOLOv8l model gave
better results in low-light conditions and images
with complex backgrounds. The YOLOv7 model,
on the other hand, performed better under standard
lighting and less complex backgrounds. This shows
that YOLOv8l is capable of better generalization in
challenging conditions.
The data augmentation techniques used affected
the performance of both models. Traditional data
augmentation techniques (e.g. rotation, translation)
applied in the YOLOv7 model increased the
generalization capability of the model. In the
YOLOv8l model, more advanced data augmentation
techniques (e.g. color space transformations,
geometric transformations) were used, which made
the model more robust to more diverse and
challenging data sets.
The comparison between the different sizes of
the YOLOv8 series (nano, small, medium, large)
shows that each model size offers different
advantages in terms of performance and
computational costs. In particular, the YOLOv8l
model, while offering high accuracy and mAP
values, requires more computational resources. On
the other hand, smaller models such as YOLOv8n
provide adequate performance with fewer
computational resources. This emphasizes the
importance of choosing the model size according to
different application requirements.
When the performance of both models is
evaluated in real-time applications, it is observed
that YOLOv8l is superior in terms of processing
speed and detection accuracy. In particular, the
YOLOv8l model offers faster detection times and
high accuracy rates, while the YOLOv7 model can
make stable detections despite its slower processing
time. These results show that model selection in
real-time applications depends not only on accuracy
rates but also on processing speed. YOLOv8l stands
out as a suitable option for applications requiring
high speed and accuracy, while YOLOv7 can be
preferred in less urgent situations and in scenarios
where stability is more important. The YOLOv7
model detected olives after 640x640 resolution, 20
batch sizes, and 100 epochs of training. The images
included situations with different densities of olives
in various environments. The model detected olives
at a wide confidence interval. In some cases, correct
detections were made with very high confidence
scores (0.9 and above), while lower confidence
scores (e.g. 0.3) were used in some detections.
Detections with low confidence scores can often
indicate false positives or detections made under
challenging conditions. This is important in
evaluating the model's balance of precision and
sensitivity, as well as its ability to generalize
different image conditions.
The YOLOv8l model was tested on similar
images at 640x640 resolution and after 50 epochs of
training. High confidence scores (0.9 and above)
were seen in most of the detections which indicated
that the model was able to make fairly accurate and
precise detections. However, the small number of
detected objects may mean that the model misses
true positives in some cases. This tests the
sensitivity of the model, especially in compelling
scenarios such as difficult lighting conditions or
overlapping objects.
When the outputs of both models were
compared, it was seen that YOLOv8l tended to
make detections with higher confidence scores. This
indicated that the model followed a more stable and
reliable learning path during training and showed a
better generalization performance. YOLOv7, on the
other hand, made detections at a wider confidence
interval, indicating that the model may make more
false positive detections in certain scenarios.
4 Discussion and Conclusion
The integration of deep learning techniques into
agriculture has attracted great attention in recent
years. Deep learning has been successfully applied
in various areas of agriculture, including smart
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
347
Volume 22, 2023
agriculture, precision agriculture, and agricultural
data analysis, [22]. In their study on image
description and deep learning in agriculture, [23]
indicated that deep learning made significant
progress in the detection of crop diseases.
Furthermore, [24] highlighted the challenges and
opportunities of applying machine and deep learning
technologies in agriculture by providing insights
regarding the specific challenges encountered in
Nigerian agriculture. The future of deep learning in
agriculture has great potential for advances in
sustainable agricultural practices, precision
agriculture, and improved decision-making. [25]
reflected the rapid development of image processing
and artificial intelligence technologies in recent
years and emphasized the critical role of deep
learning in identifying corn pests. Furthermore, the
potential of deep learning to optimize e-commerce
marketing of agricultural products was highlighted
by [26], pointing out different applications of deep
learning in various agricultural fields. [27],
identified diseases and pests in tomatoes by
applying deep learning methods to identify and
manage crop diseases in agricultural harvesting.
[28], used image segmentation of fruit
characteristics for harvest in their study using deep
learning methods in fruit recognition and evaluation
for agricultural harvesting. [29], used the YOLOv4
deep learning model to locate apples for robotic
apple harvesting. They found that the robotic
harvest detection success rate of the study was
92.9%. [30], detected bunches of grapes on the vine
using the YOLOv4 deep learning model. [31],
performed detection and classification of bell pepper
leaf diseases using artificial neural networks. With
the model they created, they found that the pepper
plant leaf was healthy or bacterial with 99.99%
predictive accuracy. [32], used deep learning
methods for robotic tomato harvesting in their study.
In addition, they determined the cut-off point for the
robotic harvesting of tomatoes by image analysis
method.
[33], conducted a prediction study by planting
seeds of 692 native genotypes in their study entitled
Fruit Yield Prediction in Pepper Using an Artificial
Neural Network. The integration of deep learning
into agriculture offers numerous opportunities for
the development of agricultural practices, improving
crop management, and advancing sustainable
agriculture. However, addressing challenges related
to data quality, model scalability, and geographic
limitations is crucial for the widespread adoption of
deep learning in agriculture. The information
provided by the reviewed academic papers provides
valuable perspectives on the current state and future
potential of deep learning in agriculture. The results
of our study are in parallel with similar research we
have stated. YOLOv7 and YOLOv8 algorithms
were used in the training of the network. A
comparison of the test performance evaluation
results of these algorithms is given in Table 1.
In Table 1, the object detection performances of
different variations of the YOLOv7 and YOLOv8l
models were compared. While the performance of
YOLOv7 after 99 epochs was notable for a
significantly lower F1 score (28.842%), each
version of the YOLOv8 series, especially YOLOv8l
(92.337% F1 score), performed much higher. These
performance measurements are critical in evaluating
the efficiency and reliability of each model in the
object detection task. The low F1 score of the
YOLOv7 model indicated that the model was not
optimal in terms of accuracy and scope of its
detections. In particular, a low Recall rate (19.56%)
indicated that the model missed a large proportion
of the objects that were actually positive. On the
other hand, YOLOv8l's high Precision (96.568%)
and very high Recall (88.462%) rates indicated that
the model showed superior performance in terms of
both accuracy and scope. YOLOv8l's
mAP@0.5:0.95 metric (94.608%) showed that the
model successfully detected even objects with high
levels of difficulty and correctly classified a wide
range of objects.
Table 1. Comparison of Model Performance
Values
Model
Epoch
F1 Score
Recall
mAP@0.5:0.95
Yolov7
99
28.842%
19.56%
62.37%
Yolov8n
49
95.293%
94.512%
87.223%
Yolov8s
49
94.469%
100.000%
90.491%
Yolov8m
49
91.772%
92.308%
90.222%
Yolov8l
49
92.337%
88.462%
94.608%
Note: The F1 Score Value in the table was calculated
according to the formula
[F1=2×(Precision.Recall/Precision+Recall].
As a result of this comparative analysis, it was
concluded that YOLOv8l had a significant
advantage in object detection performance than
YOLOv7. The high F1 score and mAP@0.5:0.95
values of YOLOv8l model showed that it was a
superior option as an advanced algorithm for object
detection, especially in various and challenging
detection scenarios. The fact that YOLOv7 showed
low performance despite a higher number of epochs,
perhaps reflected the difficulties or configuration
deficiencies that the model encountered during the
training process. The performance of YOLOv8l
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
348
Volume 22, 2023
indicated that this model should be preferred in real-
world applications as well as object detection tasks.
In our study, a comprehensive methodology and
analysis approach was made in the development and
evaluation of deep learning-based object detection
models. According to the research results, it is
understood that the advantages provided by the
YOLOv8l model will provide a basis for further
optimization and extension of the models in future
research and applications.
References:
[1] Kamaruzaman, A. S. F., Ani, A. I. C., Farid,
M., Bakar, S. J. A., Maruzuki, M. I. F.,
Setumin, S., & Hadi, M. S. (2023). Systematic
literature review: application of deep learning
processing technique for fig fruit detection
and counting. Bulletin of Electrical
Engineering and Informatics, 12(2), 1078-
1091.
[2] Yang, J., Guo, X., Li, Y., Marinello, F.,
Ercisli, S., & Zhang, Z. (2022). A survey of
few-shot learning in smart agriculture:
developments, applications, and challenges.
Plant Methods, 18(1)
[3] Wang, D., Cao, W., Zhang, F., Li, Z., Xu, S.,
& Wang, X. (2022). A review of deep
learning in multiscale agricultural sensing.
Remote Sensing, 14(3), 559.
[4] Ryo, M., Schiller, J., Stiller, S., Palacio, J. C.
R., Mengsuwan, K., Safonova, A., & Wei, Y.
(2022). Deep learning for sustainable
agriculture needs ecology and human
involvement. Journal of Sustainable
Agriculture and Environment, 2(1), 40-44.
[5] Jin, X., Zhu, X., Ji, J., Ma, Y., Xie, X., &
Zhao, B. (2023). Design and research an
online diagnosis platform for tomato seedling
facilities production diseases, Research
Square, p.1-23,
https://doi.org/10.21203/rs.3.rs-3121099/v1.
[6] Derisma, D., Rokhman, N., & Usuman, I.
(2022). Systematic review of the early
detection and classification of plant diseases
using deep learning. IOP Conference Series:
Earth and Environmental Science, 1097(1),
012042. IOP Publishing.
[7] Khairi, D. u., Ahsan, K., Badshah, G., Ali, S.
Z., Raza, S. A., Alqahtani, O., & Shiraz, M.
(2023). Comparison analysis of machine
learning classification on apple fruit quality,
Research Square, p.1-24.
https://doi.org/10.21203/rs.3.rs-3025343/v1.
[8] Habib, R. R. and Fathallah, F. A. (2012).
Migrant women farm workers in the
occupational health literature. Work, 41, 4356-
4362. IOS Press.
[9] Berg, H., Maneas, G., & Engstrom, A. (2018).
A comparison between organic and
conventional olive farming in Messenia,
Greece. Horticulturae, 4(3), 15.
[10] Tariq, S., Hakim, A., Siddiqi, A. A., & Owais,
M. (2022). An image dataset of fruitfly
species (bactrocera zonata and bactrocera
dorsalis) and automated species classification
through object detection. Data in Brief, 43,
108366.
[11] Aman Jain, Jatin Gupta, Somya Khandelwal,
Surinder Kaur. (2021). Vehicle License Plate
Recognition. Fusion: Practice and
Applications, Fusion:Practice and
Applications (FPA),4 (1), 15-21.
[12] Akiki, P. A., Akiki, P. A., Bandara, A. K., &
Yu, Y. (2020). Eud-mars: end-user
development of model-driven adaptive
robotics software systems. Science of
Computer Programming, 200, 102534.
https://doi.org 10.1016/j.scico.2020.102534
[13] Dave, K. (2018). Comparison of flow-based
versus block-based programming for naive
programmers. Toronto Metropolitan
University,Canada,Thesis.
https://doi.org/10.32920/ryerson.14652246.v2
[14] Delgado, G., Cortés, A., García, S., Loyo, E.,
Berasategi, M., & Aranjuelo, N. (2023).
Methodology for generating synthetic labeled
datasets for visual container inspection.
Transportation Research Part E: Logistics
and Transportation Review, 175, 103174.
[15] Holla, B. A., Pai, M. M. M., Verma, U., &
Pai, R. M. (2023). Enhanced vehicle re-
identification for smart city applications using
zone specific surveillance. IEEE Access, 11,
29234-29249.
[16] Dao, N. D., Le, T. V., Tran, H. T. M.,
Nguyen, Y. T. H., & Duy, T. D. (2022). The
combination of face identification and action
recognition for fall detection. Journal of
Science and Technology Issue on Information
and Communications Technology, 37-44.
[17] Ghourabi, M., Mourad-Chehade, F., &
Chkeir, A. (2023). Eye recognition by yolo for
inner canthus temperature detection in the
elderly using a transfer learning approach.
Sensors, 23(4), 1851.
[18] Lou, H., Duan, X., Guo, J., Liu, H., Bi, L., &
Chen, H. (2023). Dc-yolov8: small size object
detection algorithm based on camera sensor.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
349
Volume 22, 2023
Electronics,12(10):2323.
https://doi.org/10.3390/electronics12102323.
[19] Dehaerne, E., Dey, B., Esfandiar, H.,
Verstraete, L., Suh, H. S., Halder, S., … & De
Gendt, S. (2023). Yolov8 for defect inspection
of hexagonal directed self-assembly patterns:
a data-centric approach. 38th European Mask
and Lithography Conference (EMLC
2023),Dresden,Dermany.
[20] Asante, I., Tsun, M. T. K., Jo, H. S., &
McCarthy, C. (2023). Segmentation-based
angular position estimation algorithm for
dynamic path planning by a person-following
robot. IEEE Access, 11, 41034-41053.
[21] González-Santamarta, M. Á., Rodríguez-Lera,
F. J., & Olivera, V. M. (2023). SAILOR:
perceptual anchoring for robotic cognitive
architectures,
arXiv:2303.08204, arxiv.org.
https://doi.org/10.48550/arXiv.2303.08204.
[22] Kumar, R., Kumar, P., Aljuhani, A., Islam, A.
K. M. N., Jolfaei, A., & Garg, S. (2023). Deep
learning and smart contract-assisted secure
data sharing for iot-based intelligent
agriculture. IEEE Intelligent Systems, 38(4),
42-51.
[23] Mamat, N., Othman, M. F., Abdulghafor, R.,
Belhaouari, S. B., Normahira, M., & Hussein,
S. F. M. (2022). Advanced technology in
agriculture industry by implementing image
annotation technique and deep learning
approach: a review. Agriculture, 12(7), 1033.
[24] Umar, M.A., Sani, B.M., & Suleiman, U.
(2022). An Overview of Machine and Deep
Learning Technologies Application in
Agriculture: Opportunities and Challenges in
Nigeria. SLU Journal of Science and
Technology,4(12).
[25] Hua, L., Zhang, Y., Zhang, Y., & Zhang, D.
(2023). Maize pests identification based on
improved yolov4-tiny. Fifth International
Conference on Computer Information Science
and Artificial Intelligence (CISAI
2022),Chongqing, China.
[26] Hui, Y., Zheng, Z., & Sun, C. (2022). E-
commerce marketing optimization of
agricultural products based on deep learning
and data mining. Computational Intelligence
and Neuroscience, 2022, 1-11.
[27] Li, J. and Wang, X. (2020). Tomato diseases
and pests detection based on improved yolo
v3 convolutional neural network. Frontiers in
Plant Science, 11.
[28] Ni, X., Li, C., Jiang, H., & Takeda, F. (2020).
Deep learning image segmentation and
extraction of blueberry fruit traits associated
with harvestability and yield. Horticulture
Research, 7(1).
[29] Zhang, L., Jia, J., Gui, G., Hao, X.,Gao, W.,
Wang, M.(2018). Deep learning based
improved classification system for designing
tomato harvesting robot. IEEE Access, 6,
67940,67950.
[30] Guo, C., Zheng, S., Cheng, G., Zhang, Y.,
Ding, J. (2023). An improved yolo v4 used for
grape detection in unstructured environment.
Frontiers in Plant Science, 14.
[31] Mustafa, H., Umer, M., Hafeez, U., Hameed,
A., Sohaib, A., Ullah, S., Madni, H. A.
(2022). “Pepper Bell Leaf Disease Detection
and Classification Using Optimized
Convolutional Neural Network”, Multimedia
Tools and Applications, 82(8), 12065–12080.
[32] Bai, Y., Mao, S., Zhou, J., Zhang, B. (2023),”
Clustered Tomato Detection and Picking
Point Location Using Machine Learning-
Aided İmage Analysis for Automatic Robotic
Harvesting”, Precision Agriculture, 24(2),
727–743.
[33] Gholipoor, M., & Nadali, F. (2019). Fruit
yield prediction of pepper using artificial
neural network. Scientia Horticulturae, 250,
249–253.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
350
Volume 22, 2023
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
The authors received no financial support for their
search, authorship, and/or publication of this article.
Conflicts of Interest
On behalf of all authors, the corresponding author
states that there is no conflict of interest.
Creative Commons Attribution License 4.0
(Attribution4.0International, CCBY4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.39
Erhan Kahya, Yasin Aslan
E-ISSN: 2224-2872
351
Volume 22, 2023